PaperlayerView exact revision on arXiv

Source-linked AI summary

Protocol Compression Changes Which Party Pays: Bilateral Cost in Cross-Organization LLM Agent Communication

Janghoon Lee

arXiv:2609.06129v1cs.CL

TL;DR

Cross-organization agents pay under separate tokenizers and cost conditions, so aggregate compression savings do not establish whether both parties should adopt a protocol. The paper measures bilateral benefits through token-level comparisons and a procurement-dialogue overlay. Runtime negotiation often forms a schema but rarely settles the bargain, and its lower listed total reflects shorter sessions.

  • Problem

    Compression acceptance is unclear when each organization pays its own bill and aggregate savings may not mean both parties benefit.

    Source: §1 · p. 1

  • Method

    The paper evaluates each party’s net benefit in uncontrolled channels using preregistered token-level comparisons and a procurement-dialogue overlay.

    Source: §1 · p. 1, §3 · p. 3, §6.2 · p. 5

  • Results

    C1's listed total is 52% of C0, yet it settles the task in only 6.7% of headline dialogues.

    Source: §7.4 · p. 6

  • Takeaways & Limitations

    A pooled cost total can hide cost redistribution, while bilateral adoption depends on the sign of each party’s benefit.

    Source: §9 · p. 8, §10 · p. 8

    The bargaining evaluation covers one procurement domain with five issues, which does not represent other bargaining settings.

    Source: Limitations · p. 10

Abstract

from arXiv · show

Agents that talk across organizations exchange long messages billed by the token. A shorter notation therefore looks like a saving that costs nothing but an agreement to use it. Recent work reports the saving is conditional. Compressed notation can instead raise total tokens by 8% to 11% over a JSON baseline, when parsing failures force extra model calls. That is measured for one payer. Between two organizations neither side can install a decoder at the other end, and each pays under its own tokenizer, price, and cache state. We measure both sides. A preregistered token-level study covered 198 content-matched item pairs across six vendors, for 2,376 native-usage cells. We then overlay an English baseline, runtime schema negotiation followed by compression, and injected-schema compression on a two-party procurement bargain with an exactly enumerated feasible set. The overlay covers 1,053 completed dialogues of a 1,215-cell grid across 3 model pairs, plus a 405-dialogue rerun of the negotiated condition. Compression amplifies cross-vendor cost dispersion by a factor of 1.078, with a 95% CI of [1.066, 1.091], and two vendor pairs reverse which endpoint is cheaper. Runtime negotiation succeeds as a protocol and fails as a bargain. The parties agree a schema in 121 of 135 headline dialogues, none of them the schema we would have supplied. They settle the task in only 9 of those dialogues, and they reach impasse in 106 of them. The negotiated sessions average 10.8 turns against 17.6, and cost 52% of the English total because sessions end sooner, not because the handshake is repaid. Break-even horizons run from 20 to 70 turns, the low end only under the conditional accounting, and all of them lie above every observed English session. On one cross-vendor pair both parties keep about half their cost. On the other the receiving party pays more at a high cache-hit rate.

1 Introduction

Cross-organization compression is a bilateral economic choice because the parties bear separate costs, and aggregate savings alone do not establish that both should adopt. The paper frames adoption around each party’s net benefit and previews token-level evidence and a conversation-cost overlay.

1 Introduction

  • Separate tokenizers, bills, and repair costs make a shorter notation an economic choice for cross-organization agents.Neither endpoint necessarily controls the other’s execution environment or model.Source: §1 · p. 1
  • A protocol is mutually rational to adopt only when both parties’ net benefits are positive, even if it reduces aggregate cost.The introduction frames the object of study as a pair of benefits, one for each endpoint.Source: §1 · p. 1
  • The study uses preregistered token-level evidence for claims (i) and (ii), then overlays C0/C1/C3 to assess billed conversation cost for claim (iv).Table 1 summarizes the four claims, their evidence, status, and reporting locations.Source: §1 · p. 2, Table 1 · p. 2

2 Related Work

Prior work shows both potential savings and costs from protocol changes, but its measurements leave important parts of cross-organization, two-party adoption untested. This paper separates endpoint costs and decisions while explicitly limiting the axes its design evaluates.

2 Related Work

  • Agora reports roughly 5-fold cost reduction on single-round exchanges, while Notation Matters reports 8% to 11% more total tokens than a JSON baseline when parsing failures trigger extra inference iterations.Agora’s savings mainly come from replacing repeated LLM work with deterministic routines; its switching values are reused here as fixed demonstration thresholds, not claimed optima.Source: §2 · p. 2
  • Prior work also reports caching savings of 41% to 80%, cache-prefix instability under compression, model-level violation rates of 8% to 99%, and content-dependent token-cost rankings.These findings are reported by Lumer et al. (2026), Song (2026), He et al. (2024), Kruthof (2026), Canaverde et al. (2026), and TensorZero (2026), respectively.Source: §2 · p. 2
  • This study gives the endpoints separate tokenizers, prices, cache states, and adoption decisions, while leaving relay fidelity, fixed-protocol evaluation, and other listed axes outside its design.Prior systems measured one payer; the paper frames five questions left pooled but does not claim to answer all five.Source: §2 · p. 2

3 Scope: Uncontrolled Channels

The study concerns uncontrolled channels where at least one party cannot control key features of the other endpoint. It excludes settings where both sides can coordinate on code, grammar, tokenizer, and cache policy.

3 Scope: Uncontrolled Channels

  • An uncontrolled channel has at least one constraint: no deployment of code at the other endpoint, no enforced structured decoding there, or no advance knowledge of its model identity and settings.Cross-organization email, public APIs, and marketplace intermediaries are examples.Source: §3 · p. 2
  • The scope excludes same-organization or same-vendor systems and settings where both parties can deploy a schema before communication.Those settings can coordinate on code, grammar, tokenizer, and cache policy, unlike the bilateral incomplete-information setting studied here.Source: §3 · p. 3
  • The unit of analysis is one directed message paid for by its receiver, and the paper does not claim every cross-organization exchange should negotiate a protocol.The proposal, acceptance, rejection, or counterproposal depends on whether observable conditions justify it; negotiation can cost and fail.Source: §3 · p. 3

4 Bilateral Cost Model

The cost model keeps each endpoint’s inputs and adoption decision separate, including cache effects and negotiation or repair costs. It also treats protocol representations as potentially different by direction.

4.1 Input cost with caching

  • Each party’s input cost uses its own tokenizer, prices, cache-hit probability, and prefix length; cache-hit sweeps are limited to vendors with documented cached-input prices.The study sweeps h_i over 0, 0.5, 0.75, 0.9, and 0.99 only when that price is documented; unknown values are not imputed.Source: §4.1 · p. 3

4.2 Negotiation, compliance, and repair

  • Break-even length is party-specific, while the context-aware model sums per-turn differences to include histories, caching, and format-dependent tokenization.The model also includes expected repair cost; its negotiation-failure sweep is not reported because refusal, silence, and misunderstanding counts were not logged.Source: §4.2 · p. 3

4.3 Asymmetric protocols

  • A protocol can choose one representation per direction; symmetric representation is optimal only when it minimizes both receiver-side objectives subject to equivalent outcomes.A-to-B traffic is tokenized in B’s environment, while B-to-A traffic is tokenized in A’s.Source: §4.3 · p. 4

5 Experimental Design

The overlay compares communication conditions on the same procurement task across three model pairs, with an exactly enumerated feasible set. Negotiated conditions include the cost of negotiation, and outcome and cost records are kept distinct.

5.1 Task and endpoints

  • The overlay uses negoenv, a five-issue procurement task with 1,440 exactly enumerated allocations, across three model pairs and a 1,215-dialogue grid.The design targeted Talk is Cheap and considered TERMS-Bench as an alternative, but used negoenv as a license fallback.Source: §5.1 · p. 4

5.2 Conditions

  • C0 is the English baseline, C1 negotiates before symmetric compression, and C3 injects a schema; negotiated conditions charge both parties handshake tokens.A rejected proposal continues in English while retaining the negotiation cost.Source: §5.2 · p. 4

5.3 Outcomes

  • Each dialogue records seven cost fields per party plus a shared outcome, while compliance and repair costs are defined only when compressed turns use a known grammar.Those costs are null, not zero, for C0 and implemented C1.Source: §5.3 · pp. 4–5

6 P0: Tokenizer Divergence

The new preregistered token study separates cost-dispersion magnitude from directional reversals after the original gate failed. It finds amplified cross-vendor dispersion and two pairs whose relative token rankings reverse, while a post-hoc diagnostic distinguishes symbol families from a numeric control.

6.1 The preregistered gate failed

  • The original P0 gate failed because dispersion was 1.280 against a preregistered threshold of at least 1.300, although token rankings reversed for four vendor pairs.Both registered components were required, and the threshold and confirmatory set were not revised.Source: §6.1 · p. 5

6.2 P0b: re-operationalization and preregistration

  • P0b split the failed gate’s magnitude and sign questions into amplification ratio A and directional screen K, with separate registered pass criteria.A compares same-list maximum-to-minimum native prompt-token totals; K counts vendor pairs whose English and compressed token-order signs disagree.Source: §6.2 · p. 5

6.3 P0b results

  • 1.078 amplification (95% CI [1.066, 1.091]) and two vendor-pair reversals passed the new registered gate, with all 2,376 planned usage cells recorded.The qualifying pairs were alibaba/mistral and anthropic/google; the reporting-only spread S was 1.167 [1.146, 1.189].Source: §6.3 · p. 5

6.4 Unregistered post-hoc diagnostic: numeral density versus symbols

  • The post-hoc json_baseline control had A = 0.973 (95% CI [0.967, 0.979]), while all five symbol families exceeded its upper bound.This unregistered diagnostic supports a group-level distinction from the numeric control, not a causal decomposition of A.Source: §6.4 · p. 5

7 Results

The overlay compares billed cost, protocol agreement, and task outcomes across conditions. Negotiation often forms a schema but rarely settles the task, while cost differences remain endpoint-specific.

7.1 Headline listed cost, length, and exits (mid)

  • Implemented C1 cost 52% of C0 at h = 0 (0.00379 / 0.00735), with matched C1 − C0 total cost of −0.00355 [−0.00405, −0.00304].The archived, unimplemented C1 contrast had 0.0696 agreement and 0.930 impasse, but never ran schema negotiation.Source: §7.1 · p. 6

7.2 Protocol gate (schema_agreed)

  • Implemented C1 agreed a schema in 121/135 headline dialogues, but none matched the canonical schema and only 9/135 reached task agreement.C1 ended in impasse in 106/135 dialogues; C3 had no agreement in 112 dialogues.Source: §7.2 · p. 6

7.3 Cheaper for whom

  • On haiku-qwen, party A paid 0.201 less while B was unchanged at 0.0002, and at h = 0.9 B paid 0.072 more against A’s 0.335 less.The observed billed-cost divergence does not establish an empty mutual-benefit set because these conditions do not hold length or outcome fixed.Source: §7.3 · p. 6

7.4 Summary of the overlay

  • C1’s lower listed total reflected shorter sessions, not recovery of handshake and schema-negotiation cost within a C0-length bargain.Schema agreement therefore counted as protocol success despite the low task-settlement rate.Source: §7.4 · p. 6

8 Interpretive Break-Even Region

Section 8 estimates party-specific break-even horizons from measured costs and cache-hit-rate sweeps. Every tested horizon exceeds the observed English session length, so negotiation costs are not recouped within that session.

8 Interpretive Break-Even Region

  • The planned online policy comparison was not performed because the reduced overlay lacks its condition space and agreement is below 7% in every headline condition.Section 8 instead reports an interpretive break-even calculation on measured C0, C1, and C3 costs.Source: §8 · p. 6

8.1 Measured inputs

  • Cneg counts handshake and schema-negotiation costs, excluding compressed turns and post-failure English task turns; conditional averages include schema-agreed dialogues only.Expected Cneg averages all headline implemented C1 dialogues.Source: §8.1 · pp. 6–7
  • Figure 1 plots each party’s cost change from C0 to implemented C1; positive values indicate savings, with points at h = 0 and h = 0.9.The shaded quadrant marks cases where both parties save.Source: Figure 1 · p. 7

8.2 Party-specific horizons

  • Every tested expected and conditional break-even horizon exceeds the measured C0 mean of 17.6 turns for both parties.Table 3 reports horizons by cache-hit rate; no row is recouped at the observed English length.Source: Table 3 · p. 7, §8.2 · p. 7
  • Listed total C1 cost is below C0 because implemented sessions are shorter and usually end in impasse, not because negotiation is recouped within a C0-length bargain.The unimplemented C1 corpus had no handshake phases on the wire, so its zero Cneg reflects a runner defect.Source: §8.2 · p. 7

8.3 When compression can pay

  • The measured compression-to-English per-turn cost ratio is below 1 at every tested cache-hit rate, but each finite horizon remains beyond the observed English session.Compression can pay only if negotiation costs are amortized, both parties’ per-turn costs fall, and cache reads arrive before expiry; TTL timing was untested for the headline pair.Source: §8.3 · p. 8

9 Discussion

The discussion distinguishes cost magnitude from benefit direction and cautions that compliance and termination patterns vary with model composition.

9 Discussion

  • P0b separates compression’s cost amplification from benefit sign, because bilateral adoption depends on whether each endpoint benefits.The token-level evidence shows that a wire string can redistribute cost, but does not establish a conversation-cost mutual-benefit set.Source: §9 · p. 8
  • High C3 compliance does not explain the impasse pattern: 89 of 109 dialogues with minimum party compliance of at least 0.9 ended in impasse.This rules out low compliance as a necessary condition for the C3 fold without identifying its cause.Source: §9 · p. 8
  • Termination and schema-agreement rates vary with model-pair composition, but three pairs do not establish a model taxonomy.Direction-specific compression remains unmeasured because C2 was not run.Source: §9 · p. 8

10 Conclusion

The conclusion frames protocol adoption as bilateral and tests cost redistribution across endpoints. Runtime negotiation often agrees a schema but rarely settles the task, while compression’s cost effects depend on the pair.

10 Conclusion

  • The preregistered token-level study finds that the same wire string redistributes endpoint costs, with amplification above 1 and two vendor pairs reversing which endpoint is cheaper.The overlay tests that premise in conversation cost within channels neither party fully controls.Source: §10 · p. 8
  • Implemented runtime negotiation agrees a schema in most headline dialogues but settles the task in fewer than 7%; its lower listed total reflects shorter sessions, mostly impasses.Every interpretive horizon exceeds the measured English session length, so the handshake is not earned back within a baseline-length exchange.Source: §10 · p. 8
  • Compression’s cost effects depend on the model pair: both parties retain about half their cost on one cross-vendor pair, while the other yields a one-sided gain.At a high cache-hit rate, the receiving party on the other pair is worse off.Source: §10 · p. 8

Limitations

The study’s conclusions are bounded by incomplete coverage, a narrow task and endpoint setting, and measurement gaps that leave several protocol and policy comparisons unresolved.

Limitations

  • Cache findings are parameter sweeps, not measured TTL prevalence; most endpoints lack jointly documented numeric TTL and write-premium fields.Expired writes have no future read value by construction.Source: Limitations · p. 9
  • Complete-case bootstrap estimates may differ from full-data estimates if missingness correlates with vendor; a registered 5% missing-rate stop bounds this concern.The paired item bootstrap drops items without native usage for every vendor and surface.Source: Limitations · p. 9
  • P0b uses synthetic single-domain templates, a finite six-vendor list, and token-count reversals that do not establish conversation-cost outcomes.Generalization to naturally occurring agent traffic is untested, and dispersion extremes can change with the vendor list.Source: Limitations · pp. 9–10
  • The overlay uses one five-issue procurement domain, non-reasoning calls, and terminal reminders; other bargaining settings and reasoning-on behavior remain outside scope.The study did not test whether the same terminal outcomes appear without the reminder suffix.Source: Limitations · p. 10
  • C2, C4, and C5 were omitted, leaving direction-specific compression and handshake net-cost comparisons unmeasured; the planned switching-policy comparison was also not performed.The reduced overlay cannot compare online policies with fixed thresholds, always accept, never accept, or an oracle.Source: Limitations · p. 10
  • The final overlay stopped at its spend cap after 1,053 of 1,215 cells; the 162 unfinished shuffled-tail cells are not missing at random in calendar time.Coverage was roughly balanced across pairs and conditions.Source: Limitations · p. 10
  • C1 compliance lacks a schema-specific verifier, while refusal, silence, and misunderstanding were not logged as negotiation-failure categories.C1 compliance is recorded as compliant by construction, not as a measured perfect score.Source: Limitations · pp. 9–11
  • Retries are counted using the successful completion’s billed usage, so charges for rejected attempts are invisible; retries also concentrate on party B’s Alibaba endpoint.Most retry responses were rate limits.Source: Limitations · p. 11

Ethics Statement

This appendix documents the study’s cost-accounting assumptions, implementation and preregistration limits, screening evidence, and cache and negotiation limitations. It distinguishes cost reductions from successful task agreement, clarifies the analyses and negotiation-cost accounting, and cautions against treating token-cost orderings as judgments of model quality or broad vendor rankings.

Ethics Statement

  • The study used no human subjects, annotation, or personal data, and all measurements were paid inference on commercial endpoints within approved run caps.Recorded spend was $0.138 for P0b, $14.18 for the aborted first P1 run, $9.36 for the archived overlay, and $3.25 for the C1 rerun.Source: Ethics Statement · p. 11

A Extended Related Work

  • Adjacent work addresses relay fidelity, fixed-protocol evaluation, protocol complexity, or single-designer communication costs, rather than bilateral adoption under private costs.Neither candidate negotiation suite cleared the reuse rule, so the overlay used negoenv as a substrate, not as a contribution of the paper.Source: A Extended Related Work · p. 13, §C.4 · p. 16

B.1 Why an expired cache is not a substitute

  • When cross-organization delays exceed a numeric cache TTL, realized cache hits are zero, so cache activation and wire compression affect different token components.TTL-exceedance and negotiation-failure sweeps were defined but not reported; the headline pair had no numeric TTL to exceed, and failure categories were not logged.Source: §B.1 · pp. 13–14, §B.6 · p. 14

B.5 The primary record

  • Bilateral adoption requires both parties’ cumulative cost differences to be negative; native token counts, listed prices, cache, negotiation, and repair costs are reported separately.The decision criteria also frame the confirmatory question around whether any tested configuration benefits both parties.Source: §B.5 · p. 14, §B.4 · p. 14, §B.7 · p. 14

B.9 Deviations from preregistration

  • The P0 gate failed at 1.280 against its registered 1.300 threshold, while C2, C4, and C5 were not run and planned policy comparisons were replaced by cost arithmetic.C1’s first implementation did not negotiate and was retained only as a labelled contrast before a handshake-verified rerun.Source: §B.9 · p. 15

D.1 Per-model compressed-wire compliance

  • 6 of 6 endpoints passed the unregistered 0.5 compressed-wire screen; Mistral-small cleared it with pooled compliance of 0.544.The screen was C3 self-play, not a confirmatory test of protocol maintenance under C1 or C2.Source: §D.1 · p. 17, Table 5 · p. 17

D.3 Implications for (ii) and (iii), screening strength only

  • The screening supports treating protocol competence as a prerequisite for interpreting compressed-condition costs, but does not establish claims about asymmetric protocols or adoption.The argument applies to fixed, checkable C3 grammar; C0 has no compressed turn and C1 negotiates its grammar per dialogue.Source: §D.3 · p. 17

D.4 Cheap-pair pilot retained; excluded from the overlay

  • Mistral-small’s C3 compliance fell from 0.544 in self-play to 0.111 as party B in v4 dialogues, indicating observed multi-turn drift rather than a confirmatory result.The pilot numbers were excluded from the 1,053-dialogue overlay; Mistral was also excluded from that overlay because of 429-driven pacing and unusable compressed-cost cells.Source: §B.10 · p. 15, §D.4 · p. 17, §D.5 · p. 18

E.3 Asynchronous write loss

  • Long inter-turn delays can leave a cache write unread, but the headline OpenAI–Google implicit pair lacks documented numeric TTL and write-premium values for assigning that loss.A numeric default-path premium is documented only for Anthropic among study endpoints, so its loss was not transferred to the headline pair.Source: §E.3 · p. 19

F.2 Why the gate did not test the empty-set claim

  • P0’s maximum-to-minimum dispersion gate measured magnitude, not whether any configuration leaves both parties worse off, and depended on vendor selection and an unsupported threshold.Large dispersion does not establish an empty mutual-benefit set, and small dispersion does not rule one out; the authors call for a new registration rather than reinterpretation.Source: §F.2 · p. 20

F.3 Descriptive observations after the failed gate

  • Post-hoc P0 observations found retained-token fractions of 0.776 for DeepSeek and 0.875 for Mistral, 1.280 compressed dispersion versus 1.188 English, and four rank reversals.The 7.7% dispersion increase compares the same vendor list; rank reversals concern token counts, not net conversation cost or the empty-set claim.Source: §F.3 · p. 20

F.4 Provenance of notation-family observations

  • Round 2 symbol-dense forms had pooled dispersion 1.280, while unicode_ops reached 1.550; its Round 1 result of 1.385 gave it distinct out-of-sample provenance.The other four families exceeding 1.300 were selected after observing Round 2, so reporting them as confirmatory would incur winner’s curse.Source: §F.4 · p. 20

G.10 Numeral density versus symbols, in full

  • In an unregistered post-hoc diagnostic, json_baseline had confirmation-half A = 0.973 (95% CI [0.967, 0.979]), while all five symbol families sat above that interval.This suggests numerals alone do not explain pooled amplification on this control, but does not causally decompose digit and symbol contributions or establish family rank order.Source: §G.10 · p. 24, §G.11 · p. 24

G.12 The P0b registration in full

  • P0b preregistered amplification and sign-conflict gates on 198 new content-matched pairs, with confirmation-half family estimates and a fixed hash split.The pass rule was PASS = G1 AND G3: the 95% interval lower bound for A had to exceed 1.000, and at least one pair had to conflict in at least 95% of 10,000 bootstrap replicates.Source: §G.12 · p. 24

H.2 Schema agreement (not canonical confirmation)

  • Implemented C1 schema agreement was 0/285 for canonical matching, while schema negotiation failed in 114/405 dialogues and the handshake passed in 85/405.At h = 0, implemented headline Cneg was A 0.000513 and B 0.000816; at h = 0.9, it was A 0.000185 and B 0.000198.Source: §H.2 · p. 25, §H.3 · p. 25, §H.4 · p. 26

I.1 Search failure: wide and narrow impasse

  • The planned breakdown of handshake failures was not logged, no analytic failure-probability sensitivity was produced, and implemented headline C1 correctness was 0.281 [0.207, 0.356] (n=135).The appendix identifies Empty-F impasse as the correct conclusion and wide/narrow impasse as search_failure; C0 and C3 headline rates were 0.506 and 0.813.Source: §H.5 · p. 26, §H.6 · p. 26, §I.1 · p. 26

I.3 Compressed-wire compliance against cost

  • C3 compressed-turn compliance averaged 0.825, with a 0.429 minimum across 230 dialogues; C0 and implemented C1 had no measured violation rate.C0 has no compressed turns to check, while implemented C1 negotiates its grammar per dialogue; plotting their default 1.0 would show the instrument rather than protocol compliance.Source: Figure 2 · p. 27, §I.3 · p. 27

I.4 The analysis plan’s listed-cost cell

  • Headline schema_agreed was 0.896 versus task agreement at 0.0667; N∗ exceeded measured C0 length at every tested h, while failed-bargain compression remained cheaper on the clock.That clock-cost reduction does not mean protocol negotiation was recouped.Source: §I.4 · p. 27

I.5 Full precision

  • At h = 0, all three matched total-cost contrasts were negative: implemented C1 versus C0, C1 versus C3, and C3 versus C0.The contrasts were −0.003551 [−0.004038, −0.003021] (n=118), −0.001220 [−0.001885, −0.000558] (n=112), and −0.002369 [−0.002981, −0.001748] (n=99), respectively.Source: Table 26 · p. 28, §I.5 · p. 28

I.6 The superseded negotiation-cost numerator

  • The corrected negotiation-cost numerator is handshake plus schema-negotiation phase cost, positive in the implemented overlay and yielding N∗ above measured C0 length.The earlier C1 − C3 definition was negative on the unimplemented corpus because it mixed negotiation cost with a length difference: 5.15 turns for C1 versus 12.3 for C3.Source: §I.6 · p. 28

I.7 Compliance and impasse in the pooled C3 data

In pooled C3 dialogues, impasse rates differ across compliance bands, but pair composition and exit behavior weaken the comparison. The appendix also identifies the break-even sweep’s plotted quantities and the append-only basis for body artifacts.

I.7 Compliance and impasse in the pooled C3 data

  • 53 of 72 dialogues below 0.5 end in impasse, versus 93 of 212 at or above 0.9, but the pooled comparison does not identify a cause.The comparison uses the minimum of the two party compliance rates and pools all three model pairs.Source: §I.7 · p. 28
  • The pooled contrast mixes cross-vendor deadlock with self-play turn-cap exits, so it cannot establish a compliance effect.Of 212 high-compliance dialogues, 109 were mid and 103 nano-nano; 97 of the 103 nano-nano dialogues exited at the turn cap.Source: §I.7 · p. 28
  • Haiku-qwen C3 averages 0.375 compliance for party A and 1.000 for party B, confounding format compliance with model composition.The data cannot isolate whether failed compressed syntax blocks bargaining expressions.Source: §I.7 · p. 28

I.8 Break-even against the cache-hit sweep

  • Figure 3 plots interpretive N* against cache-hit rate h separately for parties A and B, with solid all-dialogue and dashed schema-agreed curves.The horizontal reference is the measured C0 mean of 17.6 turns; Cneg denotes handshake plus schema-negotiation phase cost.Source: Figure 3 · p. 29, §I.8 · p. 29

J Artifact Index

  • Body claims, tables, and figures are generated from append-only records and scripts, which are not recomputed or modified during manuscript assembly.The artifact index maps each body claim, table, and figure to its underlying record or script.Source: J Artifact Index · p. 30, Table 27 · p. 30
Loading 2609.06129v1…